mm/cma.c

   1 /*
   2  * Contiguous Memory Allocator
   3  *
   4  * Copyright (c) 2010-2011 by Samsung Electronics.
   5  * Copyright IBM Corporation, 2013
   6  * Copyright LG Electronics Inc., 2014
   7  * Written by:
   8  *      Marek Szyprowski <m.szyprowski@samsung.com>
   9  *      Michal Nazarewicz <mina86@mina86.com>
  10  *      Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com>
  11  *      Joonsoo Kim <iamjoonsoo.kim@lge.com>
  12  *
  13  * This program is free software; you can redistribute it and/or
  14  * modify it under the terms of the GNU General Public License as
  15  * published by the Free Software Foundation; either version 2 of the
  16  * License or (at your optional) any later version of the license.
  17  */
  18
  19 #define pr_fmt(fmt) "cma: " fmt
  20
  21 #ifdef CONFIG_CMA_DEBUG
  22 #ifndef DEBUG
  23 #  define DEBUG
  24 #endif
  25 #endif
  26 #define CREATE_TRACE_POINTS
  27
  28 #include <linux/memblock.h>
  29 #include <linux/err.h>
  30 #include <linux/mm.h>
  31 #include <linux/mutex.h>
  32 #include <linux/sizes.h>
  33 #include <linux/slab.h>
  34 #include <linux/log2.h>
  35 #include <linux/cma.h>
  36 #include <linux/highmem.h>
  37 #include <linux/io.h>
  38 #include <trace/events/cma.h>
  39
  40 #include "cma.h"
  41
  42 struct cma cma_areas[MAX_CMA_AREAS];
  43 unsigned cma_area_count;
  44 static DEFINE_MUTEX(cma_mutex);
  45
  46 phys_addr_t cma_get_base(const struct cma *cma)
  47 {
  48         return PFN_PHYS(cma->base_pfn);
  49 }
  50
  51 unsigned long cma_get_size(const struct cma *cma)
  52 {
  53         return cma->count << PAGE_SHIFT;
  54 }
  55
  56 const char *cma_get_name(const struct cma *cma)
  57 {
  58         return cma->name ? cma->name : "(undefined)";
  59 }
  60
  61 static unsigned long cma_bitmap_aligned_mask(const struct cma *cma,
  62                                              unsigned int align_order)
  63 {
  64         if (align_order <= cma->order_per_bit)
  65                 return 0;
  66         return (1UL << (align_order - cma->order_per_bit)) - 1;
  67 }
  68
  69 /*
  70  * Find the offset of the base PFN from the specified align_order.
  71  * The value returned is represented in order_per_bits.
  72  */
  73 static unsigned long cma_bitmap_aligned_offset(const struct cma *cma,
  74                                                unsigned int align_order)
  75 {
  76         return (cma->base_pfn & ((1UL << align_order) - 1))
  77                 >> cma->order_per_bit;
  78 }
  79
  80 static unsigned long cma_bitmap_pages_to_bits(const struct cma *cma,
  81                                               unsigned long pages)
  82 {
  83         return ALIGN(pages, 1UL << cma->order_per_bit) >> cma->order_per_bit;
  84 }
  85
  86 static void cma_clear_bitmap(struct cma *cma, unsigned long pfn,
  87                              unsigned int count)
  88 {
  89         unsigned long bitmap_no, bitmap_count;
  90
  91         bitmap_no = (pfn - cma->base_pfn) >> cma->order_per_bit;
  92         bitmap_count = cma_bitmap_pages_to_bits(cma, count);
  93
  94         mutex_lock(&cma->lock);
  95         bitmap_clear(cma->bitmap, bitmap_no, bitmap_count);
  96         mutex_unlock(&cma->lock);
  97 }
  98
  99 static int __init cma_activate_area(struct cma *cma)
 100 {
 101         int bitmap_size = BITS_TO_LONGS(cma_bitmap_maxno(cma)) * sizeof(long);
 102         unsigned long base_pfn = cma->base_pfn, pfn = base_pfn;
 103         unsigned i = cma->count >> pageblock_order;
 104         struct zone *zone;
 105
 106         cma->bitmap = kzalloc(bitmap_size, GFP_KERNEL);
 107
 108         if (!cma->bitmap)
 109                 return -ENOMEM;
 110
 111         WARN_ON_ONCE(!pfn_valid(pfn));
 112         zone = page_zone(pfn_to_page(pfn));
 113
 114         do {
 115                 unsigned j;
 116
 117                 base_pfn = pfn;
 118                 for (j = pageblock_nr_pages; j; --j, pfn++) {
 119                         WARN_ON_ONCE(!pfn_valid(pfn));
 120                         /*
 121                          * alloc_contig_range requires the pfn range
 122                          * specified to be in the same zone. Make this
 123                          * simple by forcing the entire CMA resv range
 124                          * to be in the same zone.
 125                          */
 126                         if (page_zone(pfn_to_page(pfn)) != zone)
 127                                 goto not_in_zone;
 128                 }
 129                 init_cma_reserved_pageblock(pfn_to_page(base_pfn));
 130         } while (--i);
 131
 132         mutex_init(&cma->lock);
 133
 134 #ifdef CONFIG_CMA_DEBUGFS
 135         INIT_HLIST_HEAD(&cma->mem_head);
 136         spin_lock_init(&cma->mem_head_lock);
 137 #endif
 138
 139         return 0;
 140
 141 not_in_zone:
 142         pr_err("CMA area %s could not be activated\n", cma->name);
 143         kfree(cma->bitmap);
 144         cma->count = 0;
 145         return -EINVAL;
 146 }
 147
 148 static int __init cma_init_reserved_areas(void)
 149 {
 150         int i;
 151
 152         for (i = 0; i < cma_area_count; i++) {
 153                 int ret = cma_activate_area(&cma_areas[i]);
 154
 155                 if (ret)
 156                         return ret;
 157         }
 158
 159         return 0;
 160 }
 161 core_initcall(cma_init_reserved_areas);
 162
 163 /**
 164  * cma_init_reserved_mem() - create custom contiguous area from reserved memory
 165  * @base: Base address of the reserved area
 166  * @size: Size of the reserved area (in bytes),
 167  * @order_per_bit: Order of pages represented by one bit on bitmap.
 168  * @res_cma: Pointer to store the created cma region.
 169  *
 170  * This function creates custom contiguous area from already reserved memory.
 171  */
 172 int __init cma_init_reserved_mem(phys_addr_t base, phys_addr_t size,
 173                                  unsigned int order_per_bit,
 174                                  const char *name,
 175                                  struct cma **res_cma)
 176 {
 177         struct cma *cma;
 178         phys_addr_t alignment;
 179
 180         /* Sanity checks */
 181         if (cma_area_count == ARRAY_SIZE(cma_areas)) {
 182                 pr_err("Not enough slots for CMA reserved regions!\n");
 183                 return -ENOSPC;
 184         }
 185
 186         if (!size || !memblock_is_region_reserved(base, size))
 187                 return -EINVAL;
 188
 189         /* ensure minimal alignment required by mm core */
 190         alignment = PAGE_SIZE <<
 191                         max_t(unsigned long, MAX_ORDER - 1, pageblock_order);
 192
 193         /* alignment should be aligned with order_per_bit */
 194         if (!IS_ALIGNED(alignment >> PAGE_SHIFT, 1 << order_per_bit))
 195                 return -EINVAL;
 196
 197         if (ALIGN(base, alignment) != base || ALIGN(size, alignment) != size)
 198                 return -EINVAL;
 199
 200         /*
 201          * Each reserved area must be initialised later, when more kernel
 202          * subsystems (like slab allocator) are available.
 203          */
 204         cma = &cma_areas[cma_area_count];
 205         if (name) {
 206                 cma->name = name;
 207         } else {
 208                 cma->name = kasprintf(GFP_KERNEL, "cma%d\n", cma_area_count);
 209                 if (!cma->name)
 210                         return -ENOMEM;
 211         }
 212         cma->base_pfn = PFN_DOWN(base);
 213         cma->count = size >> PAGE_SHIFT;
 214         cma->order_per_bit = order_per_bit;
 215         *res_cma = cma;
 216         cma_area_count++;
 217         totalcma_pages += (size / PAGE_SIZE);
 218
 219         return 0;
 220 }
 221
 222 /**
 223  * cma_declare_contiguous() - reserve custom contiguous area
 224  * @base: Base address of the reserved area optional, use 0 for any
 225  * @size: Size of the reserved area (in bytes),
 226  * @limit: End address of the reserved memory (optional, 0 for any).
 227  * @alignment: Alignment for the CMA area, should be power of 2 or zero
 228  * @order_per_bit: Order of pages represented by one bit on bitmap.
 229  * @fixed: hint about where to place the reserved area
 230  * @res_cma: Pointer to store the created cma region.
 231  *
 232  * This function reserves memory from early allocator. It should be
 233  * called by arch specific code once the early allocator (memblock or bootmem)
 234  * has been activated and all other subsystems have already allocated/reserved
 235  * memory. This function allows to create custom reserved areas.
 236  *
 237  * If @fixed is true, reserve contiguous area at exactly @base.  If false,
 238  * reserve in range from @base to @limit.
 239  */
 240 int __init cma_declare_contiguous(phys_addr_t base,
 241                         phys_addr_t size, phys_addr_t limit,
 242                         phys_addr_t alignment, unsigned int order_per_bit,
 243                         bool fixed, const char *name, struct cma **res_cma)
 244 {
 245         phys_addr_t memblock_end = memblock_end_of_DRAM();
 246         phys_addr_t highmem_start;
 247         int ret = 0;
 248
 249         /*
 250          * We can't use __pa(high_memory) directly, since high_memory
 251          * isn't a valid direct map VA, and DEBUG_VIRTUAL will (validly)
 252          * complain. Find the boundary by adding one to the last valid
 253          * address.
 254          */
 255         highmem_start = __pa(high_memory - 1) + 1;
 256         pr_debug("%s(size %pa, base %pa, limit %pa alignment %pa)\n",
 257                 __func__, &size, &base, &limit, &alignment);
 258
 259         if (cma_area_count == ARRAY_SIZE(cma_areas)) {
 260                 pr_err("Not enough slots for CMA reserved regions!\n");
 261                 return -ENOSPC;
 262         }
 263
 264         if (!size)
 265                 return -EINVAL;
 266
 267         if (alignment && !is_power_of_2(alignment))
 268                 return -EINVAL;
 269
 270         /*
 271          * Sanitise input arguments.
 272          * Pages both ends in CMA area could be merged into adjacent unmovable
 273          * migratetype page by page allocator's buddy algorithm. In the case,
 274          * you couldn't get a contiguous memory, which is not what we want.
 275          */
 276         alignment = max(alignment,  (phys_addr_t)PAGE_SIZE <<
 277                           max_t(unsigned long, MAX_ORDER - 1, pageblock_order));
 278         base = ALIGN(base, alignment);
 279         size = ALIGN(size, alignment);
 280         limit &= ~(alignment - 1);
 281
 282         if (!base)
 283                 fixed = false;
 284
 285         /* size should be aligned with order_per_bit */
 286         if (!IS_ALIGNED(size >> PAGE_SHIFT, 1 << order_per_bit))
 287                 return -EINVAL;
 288
 289         /*
 290          * If allocating at a fixed base the request region must not cross the
 291          * low/high memory boundary.
 292          */
 293         if (fixed && base < highmem_start && base + size > highmem_start) {
 294                 ret = -EINVAL;
 295                 pr_err("Region at %pa defined on low/high memory boundary (%pa)\n",
 296                         &base, &highmem_start);
 297                 goto err;
 298         }
 299
 300         /*
 301          * If the limit is unspecified or above the memblock end, its effective
 302          * value will be the memblock end. Set it explicitly to simplify further
 303          * checks.
 304          */
 305         if (limit == 0 || limit > memblock_end)
 306                 limit = memblock_end;
 307
 308         /* Reserve memory */
 309         if (fixed) {
 310                 if (memblock_is_region_reserved(base, size) ||
 311                     memblock_reserve(base, size) < 0) {
 312                         ret = -EBUSY;
 313                         goto err;
 314                 }
 315         } else {
 316                 phys_addr_t addr = 0;
 317
 318                 /*
 319                  * All pages in the reserved area must come from the same zone.
 320                  * If the requested region crosses the low/high memory boundary,
 321                  * try allocating from high memory first and fall back to low
 322                  * memory in case of failure.
 323                  */
 324                 if (base < highmem_start && limit > highmem_start) {
 325                         addr = memblock_alloc_range(size, alignment,
 326                                                     highmem_start, limit,
 327                                                     MEMBLOCK_NONE);
 328                         limit = highmem_start;
 329                 }
 330
 331                 if (!addr) {
 332                         addr = memblock_alloc_range(size, alignment, base,
 333                                                     limit,
 334                                                     MEMBLOCK_NONE);
 335                         if (!addr) {
 336                                 ret = -ENOMEM;
 337                                 goto err;
 338                         }
 339                 }
 340
 341                 /*
 342                  * kmemleak scans/reads tracked objects for pointers to other
 343                  * objects but this address isn't mapped and accessible
 344                  */
 345                 kmemleak_ignore_phys(addr);
 346                 base = addr;
 347         }
 348
 349         ret = cma_init_reserved_mem(base, size, order_per_bit, name, res_cma);
 350         if (ret)
 351                 goto err;
 352
 353         pr_info("Reserved %ld MiB at %pa\n", (unsigned long)size / SZ_1M,
 354                 &base);
 355         return 0;
 356
 357 err:
 358         pr_err("Failed to reserve %ld MiB\n", (unsigned long)size / SZ_1M);
 359         return ret;
 360 }
 361
 362 #ifdef CONFIG_CMA_DEBUG
 363 static void cma_debug_show_areas(struct cma *cma)
 364 {
 365         unsigned long next_zero_bit, next_set_bit;
 366         unsigned long start = 0;
 367         unsigned int nr_zero, nr_total = 0;
 368
 369         mutex_lock(&cma->lock);
 370         pr_info("number of available pages: ");
 371         for (;;) {
 372                 next_zero_bit = find_next_zero_bit(cma->bitmap, cma->count, start);
 373                 if (next_zero_bit >= cma->count)
 374                         break;
 375                 next_set_bit = find_next_bit(cma->bitmap, cma->count, next_zero_bit);
 376                 nr_zero = next_set_bit - next_zero_bit;
 377                 pr_cont("%s%u@%lu", nr_total ? "+" : "", nr_zero, next_zero_bit);
 378                 nr_total += nr_zero;
 379                 start = next_zero_bit + nr_zero;
 380         }
 381         pr_cont("=> %u free of %lu total pages\n", nr_total, cma->count);
 382         mutex_unlock(&cma->lock);
 383 }
 384 #else
 385 static inline void cma_debug_show_areas(struct cma *cma) { }
 386 #endif
 387
 388 /**
 389  * cma_alloc() - allocate pages from contiguous area
 390  * @cma:   Contiguous memory region for which the allocation is performed.
 391  * @count: Requested number of pages.
 392  * @align: Requested alignment of pages (in PAGE_SIZE order).
 393  *
 394  * This function allocates part of contiguous memory on specific
 395  * contiguous memory area.
 396  */
 397 struct page *cma_alloc(struct cma *cma, size_t count, unsigned int align,
 398                        gfp_t gfp_mask)
 399 {
 400         unsigned long mask, offset;
 401         unsigned long pfn = -1;
 402         unsigned long start = 0;
 403         unsigned long bitmap_maxno, bitmap_no, bitmap_count;
 404         struct page *page = NULL;
 405         int ret = -ENOMEM;
 406
 407         if (!cma || !cma->count)
 408                 return NULL;
 409
 410         pr_debug("%s(cma %p, count %zu, align %d)\n", __func__, (void *)cma,
 411                  count, align);
 412
 413         if (!count)
 414                 return NULL;
 415
 416         mask = cma_bitmap_aligned_mask(cma, align);
 417         offset = cma_bitmap_aligned_offset(cma, align);
 418         bitmap_maxno = cma_bitmap_maxno(cma);
 419         bitmap_count = cma_bitmap_pages_to_bits(cma, count);
 420
 421         if (bitmap_count > bitmap_maxno)
 422                 return NULL;
 423
 424         for (;;) {
 425                 mutex_lock(&cma->lock);
 426                 bitmap_no = bitmap_find_next_zero_area_off(cma->bitmap,
 427                                 bitmap_maxno, start, bitmap_count, mask,
 428                                 offset);
 429                 if (bitmap_no >= bitmap_maxno) {
 430                         mutex_unlock(&cma->lock);
 431                         break;
 432                 }
 433                 bitmap_set(cma->bitmap, bitmap_no, bitmap_count);
 434                 /*
 435                  * It's safe to drop the lock here. We've marked this region for
 436                  * our exclusive use. If the migration fails we will take the
 437                  * lock again and unmark it.
 438                  */
 439                 mutex_unlock(&cma->lock);
 440
 441                 pfn = cma->base_pfn + (bitmap_no << cma->order_per_bit);
 442                 mutex_lock(&cma_mutex);
 443                 ret = alloc_contig_range(pfn, pfn + count, MIGRATE_CMA,
 444                                          gfp_mask);
 445                 mutex_unlock(&cma_mutex);
 446                 if (ret == 0) {
 447                         page = pfn_to_page(pfn);
 448                         break;
 449                 }
 450
 451                 cma_clear_bitmap(cma, pfn, count);
 452                 if (ret != -EBUSY)
 453                         break;
 454
 455                 pr_debug("%s(): memory range at %p is busy, retrying\n",
 456                          __func__, pfn_to_page(pfn));
 457                 /* try again with a bit different memory target */
 458                 start = bitmap_no + mask + 1;
 459         }
 460
 461         trace_cma_alloc(pfn, page, count, align);
 462
 463         if (ret) {
 464                 pr_info("%s: alloc failed, req-size: %zu pages, ret: %d\n",
 465                         __func__, count, ret);
 466                 cma_debug_show_areas(cma);
 467         }
 468
 469         pr_debug("%s(): returned %p\n", __func__, page);
 470         return page;
 471 }
 472
 473 /**
 474  * cma_release() - release allocated pages
 475  * @cma:   Contiguous memory region for which the allocation is performed.
 476  * @pages: Allocated pages.
 477  * @count: Number of allocated pages.
 478  *
 479  * This function releases memory allocated by alloc_cma().
 480  * It returns false when provided pages do not belong to contiguous area and
 481  * true otherwise.
 482  */
 483 bool cma_release(struct cma *cma, const struct page *pages, unsigned int count)
 484 {
 485         unsigned long pfn;
 486
 487         if (!cma || !pages)
 488                 return false;
 489
 490         pr_debug("%s(page %p)\n", __func__, (void *)pages);
 491
 492         pfn = page_to_pfn(pages);
 493
 494         if (pfn < cma->base_pfn || pfn >= cma->base_pfn + cma->count)
 495                 return false;
 496
 497         VM_BUG_ON(pfn + count > cma->base_pfn + cma->count);
 498
 499         free_contig_range(pfn, count);
 500         cma_clear_bitmap(cma, pfn, count);
 501         trace_cma_release(pfn, pages, count);
 502
 503         return true;
 504 }
 505
 506 int cma_for_each_area(int (*it)(struct cma *cma, void *data), void *data)
 507 {
 508         int i;
 509
 510         for (i = 0; i < cma_area_count; i++) {
 511                 int ret = it(&cma_areas[i], data);
 512
 513                 if (ret)
 514                         return ret;
 515         }
 516
 517         return 0;
 518 }